New probabilistic transformation of imprecise belief structure

The case when the source of information provides precise belief function/mass,within the generalized power space,has been studied by many people.However,in many decision situa-tions,the precise belief structure is not always available.In this case,an interval-valued belief degree rather than a precise one may be provided.So,the probabilistic transformation of impre-cise belief function/mass in the generalized power space including Dezert-Smarandache(DSm) model from scalar transformation to sub-unitary interval transformation and,more generally,to any set of sub-unitary interval transformation is provided.Different from the existing probabilistic transformation algorithms that redistribute an ignorance mass to the singletons involved in that ignorance pro-portionally with respect to the precise belief function or probability function of singleton,the new algorithm provides an optimization idea to transform any type of imprecise belief assignment which may be represented by the union of several sub-unitary(half-) open intervals,(half-) closed intervals and/or sets of points belonging to [0,1].Numerical examples are provided to illustrate the detailed implementation process of the new probabilistic transformation approach as well as its validity and wide applicability.

New conflict representation model in generalized power space

The study on alternative combination rules in Dempster-Shafer theory (DST) when evidences are in conflict has emerged again recently as an interesting topic, especially in data/information fusion applications. The earlier researches have mainly focused on investigating the alternative which would be appropriate for the conflicting situation, under the assumption that a conflict is identified. However, the current research shows that not only the combination rule but also the classical conflict coefficient in DST are not correct to determine the conflict degree between two pieces of evidences. Most existing methods of measuring conflict do not consider the open world situation, whose frame of discernment is incomplete. To solve this problem, a new conflict representation model to determine the conflict degree between evidences is proposed in the generalized power space, which contains two parameters: the conflict distance and the conflict coefficient of inconsistent evidences. This paper argues that only when the conflict measure value in the new representation model is high, it is safe to say the evidences are in conflict. Experiments illustrate the efficiency of the proposed conflict representation model.

CLASSIFIER FUSION BASED ON EVIDENCE THEORY AND ITS APPLICATION IN FACE RECOGNITION

A multiple classifier fusion approach based on evidence combination is proposed in this paper. The individual classifier is designed based on a refined Nearest Feature Line (NFL),which is called Center-based Nearest Neighbor (CNN). CNN retains the advantages of NFL while it has relatively low computational cost. Different member classifiers are trained based on different feature spaces respectively. Corresponding mass functions can be generated based on proposed mass function determination approach. The classification decision can be made based on the combined evidence and better classification performance can be expected. Experimental results on face recognition provided verify that the new approach is rational and effective.

A novel approximation of basic probability assignment based on rank-level fusion

Belief functions theory is an important tool in the field of information fusion. However, when the cardinality of the frame of discernment becomes large, the high computational cost of evidence combination will become the bottleneck of belief functions theory in real applications. The basic probability assignment (BPA) approximations, which can reduce the complexity of the BPAs, are always used to reduce the computational cost of evidence combination. In this paper, both the cardinalities and the mass assignment values of focal elements are used as the criteria of reduction. The two criteria are jointly used by using rank-level fusion. Some experiments and related analyses are provided to illustrate and justify the proposed new BPA approximation approach.

Measuring Conflict Functions in Generalized Power Space

One of the most important open issues is that the classical conflict coefficient in D-S evidence theory (DST) cannot correctly determine the conflict degree between two pieces of evidence. This drawback greatly limits the use of DST in real application systems. Early researches mainly focused on the improvement of Dempster’s rule of combination (DRC). However, the current research shows it is very important to define new conflict coefficients to determine the conflict degree between two or more pieces of evidence. The evidential sources of information are considered in this work and the definition of a conflict measure function (CMF) is proposed for selecting some useful CMFs in the next fusion work when sources are available at each instant. Firstly, the definition and theorems of CMF are put forward. Secondly, some typical CMFs are extended and then new CMFs are put forward. Finally, experiments illustrate that the CMF based on Jousselme and its similar ones are the best suited ones.

A New Probabilistic Transformation in Generalized Power Space

The mapping from the belief to the probability domain is a controversial issue, whose original purpose is to make (hard) decision, but for contrariwise to erroneous widespread idea/claim, this is not the only interest for using such mappings nowadays. Actually the probabilistic transformations of belief mass assignments are very useful in modern multitarget multisensor tracking systems where one deals with soft decisions, especially when precise belief structures are not always available due to the existence of uncertainty in human being’s subjective judgments. Therefore, a new probabilistic transformation of interval-valued belief structure is put forward in the generalized power space, in order to build a subjective probability measure from any basic belief assignment defined on any model of the frame of discernment. Several examples are given to show how the new transformation works and we compare it to the main existing transformations proposed in the literature so far. Results are provided to illustrate the rationality and efficiency of this new proposed method making the decision problem simpler.

Modified center-based feature line classification approach

A novel classification approach called modified center-based feature line(MCFL)is proposed to reduce the computational cost of the nearest feature line(NFL)and maintain the advantages of NFL.Unlike NFL,MCFL defines a different type of feature line and utilizes both the query point’s local information and corresponding class-global information in training set.In experiments provided,the comparisons with the nearest neighbor(NN),NFL,and other NFL-refined approaches show that the computation time of MCFL can be shortened dramatically with less accuracy decreases.MCFL proposed is probably a better choice for the classification application tasks of large-scale dataset.